Optimizing AI Agents for Alpha Generation
Overview Introduces a novel systematic investment research framework for optimizing AI agents as a direct source of idiosyncratic and additive alpha. Demonstrates framework applicability on a generic fundamental equity strategy, with backtested results showing that post-optimized agents can achieve higher, statistically significant alpha than a range of generic agents and models. Provides a detailed framework specification for industry evaluation and replicability. Key findings AI agent stochasticity is real and impactful. Due to implementation constraints, virtually all LLMs are non-deterministic. Agentic capabilities further amplify this stochasticity and can significantly affect investment outcomes; under identical conditions, an OpenClaw agent could vary over 4× raw output and 1.5× in final performance than a single-turn agent. A single backtest from your AI agent is just one draw from a distribution, not true skill. Nonetheless, outputs are mostly noise around a persistent character that can be modeled through repeated runs and ensembling. More agentic is not better — heck, even a better model is not better! Agent performance in real financial settings can diverge from conventional understanding and benchmarks. In our study on idiosyncratic fundamental skills, single-turn agents (floor LLMs) consistently outperform more sophisticated open-source variants on performance metrics, by more than 30-100% after factor controls. Within floor LLMs, more intelligent but hallucination-prone models can also underperform peers; Grok 4.5 underperforms Grok 4.3, having twice the hallucination rate. Furthermore, agentic components do not interact monotonically — a model that performs better with one scaffold may not with another. There is an explanation for the findings above: error rate. By understanding the mechanics, we can configure an ex-ante optimal agent and improve iteratively. Net performance gain ≈ capability gain − (base error × amplification). Using this principle, we first-shot an ex-ante agent that outperformed all previous configurations and became the only one to clear a credible significance bar (NW t > 3) for idiosyncratic IC. Practical learnings: memory, reflection, and debate often introduce more noise and error than insight; a grounded mid-tier model beat a “smarter” one that hallucinates; split sub-agent roles by information stream; verify claims against the raw data; prefetch standard data packs instead of open-ended tool loops when possible… See results here. ...